AI & Machine Learning
foo bar
We'll help you leverage cloud services to predict outcomes and make informed decisions with AI and machine learning.
Frequently asked questions
Machine learning is a branch of artificial intelligence that makes it possible for computers to learn from data without being explicitly programmed. It's a field of science that's rapidly growing, and it gives businesses the power to make highly accurate predictions about everything from sales to employee productivity.
The key to machine learning is representation. Machine learning algorithms can automatically create models based on data. These models can then be used to make predictions about future events or identify previously unseen elements of the world.
In machine learning terminology, a model can be seen as a function or class that is trained to solve a particular task. Just as with regular functions in your programming language, it accepts one or more inputs such as text, image, tabular data, etc., and provides one or more outputs, also called predictions. For example, the model can be categorical, meaning it predicts one of the pre-defined labels (classes) such as "positive review" or "negative review", or a regression model that predicts values such as Bitcoin price in the next 5 minutes.
In classical software engineering, you define the function and create its body that will process the input in a specific way, providing well-defined output. In machine learning this process is reversed; you provide the model with inputs and outputs, and it will iteratively learn and improve on how to process the inputs to select the correct outputs. This is called training and after it's done, the model can reason even over the previously unseen data.
It heavily depends on your task. With modern approaches like transfer learning, sometimes we can create production-ready models with just a few dozen samples. But generally speaking, the more data you can provide us the better, and its quality plays a very important role. The higher the quality of the input data, the better the results.
The exact cost to run your models will depend mostly on the amount of data you need to process. If it's just a few hundred samples per day, an instance with 1-2 vCPUs running a few hours per day will probably be enough for you. If you need to do inference on millions of messages or images per month, you will probably need a GPU instance to handle that amount of data in a reasonable time, which generally costs more.
Today, successful applications of machine learning are all around us. Google is using language models to improve its search results. Alexa is using speech models to understand you better. Tesla is using state of the art AI to see the world around and make instant decisions. However, thanks to open research, every company, no matter how big or small, can benefit from machine learning. You could benefit from use-cases as: estimating the revenue of your clients, categorizing data into predefined classes, recommendation systems to increase your sales, planning the most effective transportation routes, and more!
Generally speaking, you don't need to worry that such information will be leaked by the model in any way. If the model is trained to predict image class as "dog" or "cat", it's impossible for it to reveal any hidden information to the end-users. However, if you are worried about distributing such data, we can empower a technique called data masking or ppi masking, which will hide any secret information before passing data to the researchers.
You don't need any previous experience with AI or ML. We will help you with everything you need and in the end, every software engineer will be able to operate your models.
For any task you want to solve, we will define core metrics – values with which we can track how well the model is doing. For instance, if the task is to predict stock value, the metric can be expressed as a difference between predicted and real value, in this scenario we would aim to minimize it.
Machine learning is all about learning how to make decisions based on data. Imagine you would like to create an email spam filter. After a while, you may notice that most spam emails contain inappropriate words that try to attract users for instant profits or other stuff. So you create a simple function that will mark as spam every email that contains such a word. But after a while, your users start to complain that even regular emails are being marked as spam. So you improve your solution and add more complex rules, perhaps some combination of allowed words. Improving and evaluating your function after every feedback. You can see that creating such a function would be a very tedious and complex task, that's where machine learning can save you! Instead of defining these rules yourself, you provide your models with data samples, mapping each message to a label "SPAM" or "OK" and using the training procedure, the model will infer these complex rules for you.
Machine learning is a great technique to find complex relationships in your data to solve well-defined problems. Whenever your inputs are tables, images, texts or almost anything else, as long as you can define the desired output, there is a good chance that the machine learning model can learn how to achieve it. Contact us and we can have a consultation about your use case!
After selecting the best model, we will evaluate its performance in production on unseen data. To be absolutely sure the model behaves as you desire, after some time we will also collect feedback on the model's predictions in production and evaluate it on the selected metrics.
Deep learning is a subset of machine learning. Generally speaking, you are training any model on your data, you are doing machine learning, However, some of these models can be from the category of deep learning. In machine learning models, you usually have to create inputs (also called features) manually, this can be for example number of occurrences of word X or Y, the length of the text, the number of capital letters, or anything else you can think of. For some tasks, such as predictions on tabular data from the database, this is completely fine as features like this are already available. However, you can see that creating such features manually is very hard for texts and images and that's where deep learning excels. Deep learning models are able to automatically learn how to extract features from raw inputs such as images, texts, audio and many others and afterwards, it can learn how to solve a particular task. Nowadays, deep learning dominates in the majority of tasks and it's usually a go-to solution for many of them.
